CzeSL – a Learner Corpus of Czech

• The Corpus of Czech as a Second Language
• A part of the AKCES/CLAC project (the Czech Language Acquisition Corpora)
• For the official site see AKCES – Akviziční korpusy českého jazyka (in Czech)
• An outdated site: Investice do rozvoje vzdělávání v oboru čeština jako cizí jazyk (Investments into Teaching Czech as a Foreign Language)
• Funded since 2009 from several projects:
  • 2009–2012: European Social Funds (ESF) – Innovative approach to teaching Czech as a second language, no. CZ.1.07/2.2.00/07.0119
  • 2012–2016: Ministry of Education, Youth and Sports – Czech National Corpus, no. LM2011023
  • 2016–2018 (extended to mid-2020): Grant Agency of the Czech Republic – Non-native Czech from the Theoretical and Computational Perspective, no. 16-10185S
• Alternative address of this site: http://utkl.ff.cuni.cz/learncorp/

• Tags: F – formal, G – grammar-based, MD – multi-dimensional, I – implicit
• TH (target hypothesis): K – correction suggested by the proofing tool, 2T – successive corrections in the 2T scheme, T2 – correction at Tier 2, 2T+ – more than 2 successive corrections
• Linguistic annotation: M – morphology (lemmas and morphosyntactic tags), S – syntax (structure and functions)
• Access: S – searchable on-line, D – downloadable in full as a dataset
• Year: year of the first release

• 12.4 thousand transcribed texts, 2 mil. words, 2.5 mil. tokens
• Plain = without annotation, without metadata
• Consists of 3 parts:
  • Texts written by foreign learners of Czech (ciz): 8,109 texts, 1,161 thousand tokens
  • Academic texts written by foreign students in Czech (kval): 174 texts, 732 thousand tokens
  • Texts written by Czech students with Romani background (rom), i.e. an ethnolect of Czech: 4,105 texts, 428 thousand tokens
• Searchable from the Czech National Corpus site
• Downloadable from the LINDAT-Clarin data repository as two subcorpora:
  • AKCES 3 – includes texts produced by non-native students of Czech
  • AKCES 4 – includes texts produced by students growing up in socially excluded communities
• Description in English and Czech

• Czech as a Second Language with Spelling, Grammar and Tags
• 8,617 transcribed texts, 111 thousand sentences, 1 mil. words, 1.1 mil. tokens
• Extends the “foreign” (ciz) part of CzeSL-plain by texts collected in 2013
• Original forms and automatic corrections are tagged, lemmatized and assigned error labels
• Most texts have metadata attributes (30 items) about the author and the text
• Searchable from the Czech National Corpus site, metadata available from this site are in Czech
• Dowloadable as AKCES 5 (CzeSL-SGT) Release 2 with the metadata in English. The original release is still available from AKCES 5 (CzeSL-SGT). We suggest the use of Release 2, where a number of bugs were fixed, including an issue in the metadata: native speakers of Ukrainian were labelled as speakers of “another Indo-European language” rather than as speakers of a Slavic language. Release 2 is now also a validated XML document with all annotation represented as XML attributes.
• Description in English and Czech

• Includes texts collected, transcribed and manually annotated within the ESF project, see a description in English.
• Annotation manual in Czech
• Transcription manual in Czech
• Appendix to Transcription manual in Czech
• Transcription manual – summary in English; please note that formatting specifics concerning the markup of manuscripts in the transcription manual, its appendix and the summary are outdated.

• Includes subsets of the ciz and rom parts of CzeSL-plain, i.e. the manually annotated Czech texts written by non-native learners and by speakers of the Roma ethnolect of Czech, the total of about 330 thousand tokens.
• Texts of about 208 thousand tokens are annotated independently by two annotators.
• Texts are searchable on-line via the SeLaQ tool, using the CzeSL-native format:
  • SeLaQ is a purpose-built corpus manager. See its menu for instructions.
  • Note that metadata and graphical display of links between annotated word tokens are not available in SeLaQ.
• Downloadable from here in the feat format, with metadata

• CzeSL-man v1 contains 645 texts (128 thousand tokens) from CzeSL-SGT, including 298 doubly annotated texts (59 thousand doubly annotated tokens).
• Most texts are equipped with metadata about the author, the text and the annotation process.
• CzeSL-man can be searched or downloaded. Although the set of texts in the searchable and the downloadable versions are identical, they differ in how they represent the annotation.
• CzeSL-man v1 downloadable:
  • Downloadable from https://bitbucket.org/czesl/czesl-man/
  • This release is in the PML format, generated by the feat tool.
  • Each text with its annotation consists of several related files.
  • Some of the texts are independently annotated twice.
• CzeSL-man v1 searchable:
  • Searchable by KonText: https://kontext.korpus.cz/first_form?corpname=czesl-man
  • Differs from both CzeSL-man v0 and CzeSL-man v1 downloadable in two aspects:
    • There are no texts with alternative error annotation: each text is annotated by a single annotator
    • The two-tier annotation scheme is radically modified to fit the token-based setup of the search tool.
  • Apart from that, the content and metadata are identical to CzeSL-man v1 downloadable and the search options to those of CzeSL-SGT.
  • The main feature in the annotation is the reversal of the source text and its annotation. The correction at Tier 2 is assumed to be the basis for the annotation. The tokens of this corpus represent the words at Tier 2. The original text is added as annotation of the Tier 2 tokens. Each token of the corrected text receives its corresponding Tier 0 form and a Tier 2 error label as attributes.
  • This annotation discards any Tier 1 corrections and error tags, and simplifies other than 1:1 links between tokens at Tier 0 and Tier 2.
  • Tier 2 is parsed in a way similar to some other Czech corpora searchable in KonText, such as SYN2015.

• In this release the two-tier error annotation is represented as pairs of XML elements err and corr. An ill-formed portion of the source text is enclosed within the err structure, immediately followed by its correction, enclosed within the corr structure.
• Apart from the error annotation, the content and metadata are the same as in CzeSL-man v1.
• Linguistic annotation (tags and lemmas) is provided for all tokens at Tier 0 and Tier 2.

• Includes a subset of CzeSL-SGT, hand-corrected, but not error-tagged, in 2017–2018, according to the 2T scheme.
• The corpus includes about 1300 texts (180 thousand tokens), selected from those that had not been manually error-annotated before (are not part of CzeSL-man).
• The selection was meant to make the manually annotated part of CzeSL more balanced in terms of L1 and CEFR level.
• Downloadable from here in the feat format

• Includes a subset of texts from CzeSL-man, semi-automatically annotated by a multi-domain tagset with a focus on morphology, see the annotation manual (in Czech)
• The dataset is available from https://bitbucket.org/czesl/czesl-md in the brat format
• The texts can also be viewed and searched here using brat
• To be extended to all CzeSL-man texts

• Texts from CzeSL-man with syntactic annotation according to the Universal Dependencies (UD) standard
• Available from the LINDAT/CLARIN repository
• CzeSL-man, CzeSL-MD and CzeSL-UD will eventually be merged into a single corpus with multiple types of annotation

• CzeSL Grammatical Error Correction Dataset – manually annotated texts, converted to a format intended for NLP applications
• Sentence pairs of original and corrected versions of Czech sentences collected from essays written by both non-native learners of Czech (CzeSL-man) and Czech pupils with Romani background
• Downloadable from http://hdl.handle.net/11234/1-2143
• See also AKCES-GEC

• AKCES Grammatical Error Correction Dataset for Czech – a grammar error correction corpus for Czech generated from a subset of AKCES. It contains train, dev and test files annotated in M2 format. Note that in comparison to CZESL-GEC dataset, this dataset contains separated edits together with their type annotations in M2 format and also has twice as many sentences.
• Downloadable from http://hdl.handle.net/11234/1-3057
• See https://arxiv.org/pdf/1910.00353.pdf for a detailed description

• Work in progress: will eventually include all available Czech texts written by non-native learners
• See CzeSL in TEITOK at the ICTL site

• Annotation editor feat, used for multi-level manual error annotation of CzeSL-man
• Annotation editor brat, used for multi-domain error annotation of CzeSL-MD
• Tagger and lemmatizer of Czech Morphodita, used for morphological annotation of CzeSL-SGT
• Spelling/grammar checker Korektor, used for automatic correction of CzeSL-SGT
• Error identifier (see Jelínek, 2017), used for automatic identification of some types of errors in CzeSL-SGT
• Multi-level concordancer SeLaQ, used for basic searching in CzeSL-man
• Standard concordancer Manatee/KonText, used for searching in CzeSL-plain and CzeSL-SGT
• General corpus tool TEITOK, currently used for building, editing and viewing learner corpora hosted by the Institute of Theoretical and Computational linguistics (see Learner corpora at ICTL)

Bibliography

NEW:

Rosen, A., Hana, J., Hladká, B., Jelínek, T., Škodová, S., and Štindlová, B. (2020). Compiling and annotating a learner corpus for a morphologically rich language – CzeSL, a corpus of non-native Czech. Karolinum, Charles University Press, Praha. Print copy, e-book CU Digital Repository